Results!

Year over Year stability of Separation Differential

Below is a table showing how correlated a receivers average Separation Differential is year-over-year. Turns out, among receievers with at least 70 targets in each of the last 3 years, there is a low, but non-zero correlation between those figures over three seasons. Considering the variable nature of separation, this is to be expected, but I would be worth looking into why these correlation values aren't more stable year over year.

How do the players stack up?

Top 10 Highest Average Separation Above Expectation since 2017

Top 10 Lowest Average Separation Above Expectation

2019 Only

Exceptional Separation Rate

'Exceptional Separation Rate' tallies all the instances where a receiver exceeds their predicted separation, and divides that by their total targets to see how often they are exceeding expectations. This metric will be less prone to outliers.

Why should you care?

The goal of this exercise is try and extract information from recevier separation, and the analysis doesn't end with these numbers. In an attmept to see how this stat looks on the team scale, I correlated win percentage over the last three seasons with the teams average team separation differential to see how this stat did to predict team success.

Amazingly enough, teams with higher average separation differentials won more often than teams with lower separation differentials.

I know what you are thinking,

"of course it is! receiver separation is a good thing, so teams with higher separation win more!"

Yes, that is true, BUT, separation differential is actually a better predictor of team success than average separation.

Player Evaluation

The interactive plot above shows the differences between predicted and actual average separations of receivers with at least 175 targets in the last three seasons.

Players above the diagonal line excel at gaining separation relative to their peers.

This could be a separate clustering project altogether, but you will notice 4 distinct groups of players as you travel up and to the right along the line.

  1. Bottom Left
    • Tall, Iso wide receviers that go deep often and operate in tight windows
      • Best Separation Differential: DeAndre Hopkins, John Brown
      • Worst Separation Differential: Kenny Golladay, Allen Robinson
  2. Middle-Left
    • Average-sized receivers that move around (between wide, slot) more often and run a larger variety of routes
      • Best: Davante Adams, Sammy Watkins
      • Worst: Demaryius Thomas, Rob Gronkowski
  3. Middle-Right
    • These receivers work almost exclusivly from the slot, and most tend to run shallower routes
      • Best: Tyreek Hill, Cooper Kupp
      • Worst: Larry Fitzgerald, Golden Tate
  4. Top Right
    • Mostly Tight Ends and and exclusive Slot receivers
      • Best: Jack Doyle, Austin Hooper
      • Worst: Jamison Crowder, Kyle Rudolph

Main Takeaways

  1. Separation Above Expectation is a better predictor of team success than Separation
    • Separation Above Expectation has a 0.66 correlation to Win % vs .48 for Avg Separation
  2. Receiver Separation must be compared holistically by controlling for factors like depth of target, pre-snap cushion, and nearest defender
    • Tyreek Hill, Davante Adams, Cooper Kupp, and Tyler Lockett some of the best route-runners/separation-getters in the NFL.
  3. Compare receviers within clusters, and select for positive Separation Above Expectation
    • From a team-building perspective, you cannot replace Antonio Brown with JuJu Smith-Schuster (as we all saw). Players with different body types and abilities have different team roles, and need to be accounted for when building a receiving corps.

Potential Improvements

If I had more time/resources/data, what would I do to make these predictions better?

  1. New Feature - Absolute Yardline
    • Absolute yardline (where the play starts from on the field) should be helpful
    • What type of coverage is the defense running? Man v Zone?
  2. New Feature - Time Left + Score Differential
    • What is the game situation? Is one team down 21 points and playing catch up? Is it tied in the 4th quarter?
  3. Create different model for each position
    • Each position is very specialized, that's the beauty of football
      • This model might not understand all of the nuances of separation for every position, but it can be generalized to be applied to all (non-backfield) receivers. There is always a trade-off between applicability and interpretability. If I build separate models for WRs/TEs/RBs it may be more insightful for each position group, but also inherently more complicated to understand and compare accross positions.
  4. Hyperparameter tuning
    • In the Light Gradient Boosted Model, hyperparameter tuning is time-consuming and computationally heavy. With more time, I would spend more time fine-tuning these paramters to create the most accurate model
  5. Targets only
    • This dataset only includes instances when the player was targeted, so it has limitations. If we wanted to include all routes, we would have to create a proxy for separation (perhaps at pass forward).

Thanks for reading!

Thank you all for reading. If you have any questions, suggestions, or any feedback at all, feel free to reach out at JesseDCohen@gmail.com.